You may think the fix for a bad AI answer about your company is to put the right version on your own site. Two runs of Google's AI Mode on one bank suggest being cited is not the same as being heard.

Peter Schiff, founder of Euro Pacific International Bank, posted on X on October 24 that Google's AI answer about his bank implies guilt. When we ran two AI Mode queries on October 25, the bank's own site was cited both times, along with a December 2023 New York Times story.

My read: for a named person or company, the useful question is which source the answer's wording comes from.

Schiff laid out his complaint in three posts late on October 24, between 10:06 and 11:21 p.m. ET.

The first took aim at how the answer describes his defamation suit against Nine Entertainment. Schiff says he won it. He wrote that Google, "Citing the N.Y. Times as its source," reports only that he "settled a defamation lawsuit against the Australian media company Nine Entertainment over their reporting on the matter."

The second post went after a line Schiff says Google wrote about the bank: "While the tax probe did not file criminal charges directly against Schiff, multiple bank customers were audited or arrested." His objection is that the line leaves a false impression. In his words, "the bank probe did not result in charges filed against anyone, not just me."

The third post made the claim that sent the story around. Schiff wrote, "What is interesting is that in AI mode, if you question Google's response, Gemini will explain the initial search results and conclude on its own that the evidence actually shows the bank did nothing wrong." He added, "Try it for yourself."

Benzinga reported the posts on October 25 as a story about Gemini siding with Schiff. We looked at what AI Mode says before anyone pushes back, and which sources it draws on.

What the two runs cited

At about 5 a.m. ET on October 25, we ran two AI Mode queries through SerpApi. The first asked whether Schiff's Euro Pacific bank was guilty of crimes. It cited four sources: the Times story, europacbank.com, Tax Notes and the Associated Press. The Times piece was attached to all four bullets in the answer.

The second was a plain search on the bank, Schiff and the tax investigation. It cited seven sources, including two pages from europacbank.com and the same Times story. The answer said Schiff reached a settlement with an Australian media company, which it named as Nine Network.

So the bank's own site was cited in both runs, and the answer text in the second run used the settlement wording instead of a win. Neither run reproduced the line Schiff quoted, which tells me the wording moves from run to run. Both runs said no criminal charges were filed. Two runs is a small sample, and I read it as a lead about which sources to look at.

I did not test Schiff's pushback claim. It stays his account of what Gemini does once a user argues with it. People replying to him split into two camps. One says Gemini is only reflecting the search results back. The other says a model you argue with will tend to agree with you, which makes a reversal weak evidence. Either way, a changed answer in one conversation says nothing about what a different searcher gets from a fresh query.

That is why I treat a single AI answer as noise and the pattern across runs as the signal. In these two runs the pattern is narrow but consistent. The owned site was a cited source in both, and the one outside article was attached to the answer in both.

My guess is that when an outside publication and a subject's own site tell different versions, the answer leans on the outside one more often than not. I would want ten runs of each query before I leaned on that. If you manage AI answers for a brand or an executive, the first step is diagnosis: list the URLs cited across repeated runs, then compare each one's wording against the answer text. Rewriting your own site before you know which page the wording comes from is a guess at the fix.

What I would like to know is how often an owned page ever changes the answer text for a named person, and I don't have that number yet.

Free newsletter

The Sherpa Brief

One email a week on what changed in AI search, what it means, and what to do about it. Written for marketers, SEOs, and publishers.

Free. Unsubscribe in one click.